/sgmse

Score-based Generative Models (Diffusion Models) for Speech Enhancement and Dereverberation

Primary LanguagePythonMIT LicenseMIT

Speech Enhancement and Dereverberation with Diffusion-based Generative Models

Diffusion process on a spectrogram: In the forward process noise is gradually added to the clean speech spectrogram x0, while the reverse process learns to generate clean speech in an iterative fashion starting from the corrupted signal xT.

This repository contains the official PyTorch implementations for the papers:

Audio examples and supplementary materials are available on our SGMSE project page and EARS project page.

Follow-up work

Please also check out our follow-up work with code available:

Installation

  • Create a new virtual environment with Python 3.11 (we have not tested other Python versions, but they may work).
  • Install the package dependencies via pip install -r requirements.txt.
    • Let pip resolve the dependencies for you. If you encounter any issues, please check requirements_version.txt for the exact versions we used.
  • If using W&B logging (default):
    • Set up a wandb.ai account
    • Log in via wandb login before running our code.
  • If not using W&B logging:
    • Pass the option --nolog to train.py.
    • Your logs will be stored as local CSVLogger logs in lightning_logs/.

Pretrained checkpoints

  • For the speech enhancement task, we offer pretrained checkpoints for models that have been trained on the VoiceBank-DEMAND and WSJ0-CHiME3 datasets, as described in our journal paper [2]. You can download them here.
    • SGMSE+ trained on VoiceBank-DEMAND: gdown 1_H3EXvhcYBhOZ9QNUcD5VZHc6ktrRbwQ
    • SGMSE+ trained on WSJ0-CHiME3: gdown 16K4DUdpmLhDNC7pJhBBc08pkSIn_yMPi
  • For the dereverberation task, we offer a checkpoint trained on our WSJ0-REVERB dataset. You can download it here.
    • SGMSE+ trained on WSJ0-REVERB: gdown 1eiOy0VjHh9V9ZUFTxu1Pq2w19izl9ejD
    • Note that this checkpoint works better with sampler settings --N 50 --snr 0.33.
  • For 48 kHz models [3], we offer pretrained checkpoints for speech enhancement, trained on the EARS-WHAM dataset, and for dereverberation, trained on the EARS-Reverb dataset. You can download them here.
    • SGMSE+ trained on EARS-WHAM: gdown 1t_DLLk8iPH6nj8M5wGeOP3jFPaz3i7K5
    • SGMSE+ trained on EARS-Reverb: gdown 1PunXuLbuyGkknQCn_y-RCV2dTZBhyE3V

Usage:

  • For resuming training, you can use the --ckpt option of train.py.
  • For evaluating these checkpoints, use the --ckpt option of enhancement.py (see section Evaluation below).

Training

Training is done by executing train.py. A minimal running example with default settings (as in our paper [2]) can be run with

python train.py --base_dir <your_base_dir>

where your_base_dir should be a path to a folder containing subdirectories train/ and valid/ (optionally test/ as well). Each subdirectory must itself have two subdirectories clean/ and noisy/, with the same filenames present in both. We currently only support training with .wav files.

To see all available training options, run python train.py --help. Note that the available options for the SDE and the backbone network change depending on which SDE and backbone you use. These can be set through the --sde and --backbone options.

Note:

  • Our journal preprint [2] uses --backbone ncsnpp.
  • For the 48 kHz model [3], use --backbone ncsnpp_48k --n_fft 1534 --hop_length 384 --spec_factor 0.065 --spec_abs_exponent 0.667 --sigma-min 0.1 --sigma-max 1.0 --theta 2.0
  • Our Interspeech paper [1] uses --backbone dcunet. You need to pass --n_fft 512 to make it work.
    • Also note that the default parameters for the spectrogram transformation in this repository are slightly different from the ones listed in the first (Interspeech) paper (--spec_factor 0.15 rather than --spec_factor 0.333), but we've found the value in this repository to generally perform better for both models [1] and [2].

Evaluation

To evaluate on a test set, run

python enhancement.py --test_dir <your_test_dir> --enhanced_dir <your_enhanced_dir> --ckpt <path_to_model_checkpoint>

to generate the enhanced .wav files, and subsequently run

python calc_metrics.py --test_dir <your_test_dir> --enhanced_dir <your_enhanced_dir>

to calculate and output the instrumental metrics.

Both scripts should receive the same --test_dir and --enhanced_dir parameters. The --cpkt parameter of enhancement.py should be the path to a trained model checkpoint, as stored by the logger in logs/.

Citations / References

We kindly ask you to cite our papers in your publication when using any of our research or code:

@inproceedings{welker22speech,
  author={Simon Welker and Julius Richter and Timo Gerkmann},
  title={Speech Enhancement with Score-Based Generative Models in the Complex {STFT} Domain},
  year={2022},
  booktitle={Proc. Interspeech 2022},
  pages={2928--2932},
  doi={10.21437/Interspeech.2022-10653}
}
@article{richter2023speech,
  title={Speech Enhancement and Dereverberation with Diffusion-based Generative Models},
  author={Richter, Julius and Welker, Simon and Lemercier, Jean-Marie and Lay, Bunlong and Gerkmann, Timo},
  journal={IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume={31},
  pages={2351-2364},
  year={2023},
  doi={10.1109/TASLP.2023.3285241}
}
@inproceedings{richter2024ears,
  title={{EARS}: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation},
  author={Richter, Julius and Wu, Yi-Chiao and Krenn, Steven and Welker, Simon and Lay, Bunlong and Watanabe, Shinjii and Richard, Alexander and Gerkmann, Timo},
  booktitle={ISCA Interspeech},
  year={2024}
}

[1] Simon Welker, Julius Richter, Timo Gerkmann. "Speech Enhancement with Score-Based Generative Models in the Complex STFT Domain", ISCA Interspeech, Incheon, Korea, Sep. 2022.

[2] Julius Richter, Simon Welker, Jean-Marie Lemercier, Bunlong Lay, Timo Gerkmann. "Speech Enhancement and Dereverberation with Diffusion-Based Generative Models", IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2351-2364, 2023.

[3] Julius Richter, Yi-Chiao Wu, Steven Krenn, Simon Welker, Bunlong Lay, Shinji Watanabe, Alexander Richard, Timo Gerkmann. "EARS: An Anechoic Fullband Speech Dataset Benchmarked for Speech Enhancement and Dereverberation", ISCA Interspeech, Kos, Greece, 2024.